基于混合深度神经网络的燃料电池混合动力汽车能量管理优化研究

Energy management optimization of fuel cell hybrid electric vehicles based on hybrid deep neural networks

为提升燃料电池混合动力汽车在短程驾驶过程中的燃油经济性,文章构建了一种基于混合深度神经网络的车速预测模型VBS-Net。该模型不仅进一步改良了基于VGG-Net结构的卷积网络,还引入了双向长短期记忆神经网络,对整个车速预测序列的时空依赖关系进行有效学习。同时考虑预测时域和输入序列长度对短程车速预测精度的影响,利用贝叶斯优化超参数进一步提升VBS-Net的预测精度。为解决能量管理策略的在线优化和计算效率问题,设计了一种基于多目标优化的模型预测控制(MPC)能量管理策略。该策略可以实现对氢气消耗、锂电池充电状态(SOC)维持、燃料电池使用效率三者的平衡优化。最后在3种实车工况下,将所提策略与基于规则的策略相比,燃油经济性分别提升了7.25%,9.94%和19.23%,且有更好的SOC维持特性。

In order to improve the fuel economy of fuel cell hybrid electric vehicles during short range driving,a vehicle speed prediction model structure VBS-net based on hybrid deep neural network was constructed.This structure not only further improves the convolutional network based on the VGG-Net structure,but also introduces a bidirectional long short-term memory neural network to effectively learn the spatiotemporal dependencies of the entire vehicle speed prediction sequence.Simultaneously considering the influence of prediction time domain and input sequence length on the prediction accuracy of short-range vehicle speed problems,Bayesian optimization hyperparameters are used to further improve the prediction accuracy of VBS-Net.To address the online optimization and computational efficiency issues of energy management strategies,a multi-objective optimization based on model predictive control(MPC)energy management strategy was designed.This strategy can achieve a balance and optimization of hydrogen consumption,lithium battery state of charge(SOC)maintenance,and fuel cell utilization efficiency.Finally,under actual vehicle conditions,the proposed strategy was compared with rule-based strategies,resulting in fuel economy improvements of 7.25%,9.94%and 19.23%,and better SOC maintenance characteristics.